Software Engineering Courses (SECourses)•2y ago

Any tips for cartoon character lora training? Which caption approach and which caption tool would yo

Any tips for cartoon character lora training?
Which caption approach and which caption tool would you advise?
Running locally on 16G vram, 4080. Using Kohya.
@Furkan Gözükara SECourses no rush, tagging for visibility.

JP525•7/31/24, 7:09 AM

Hi everyone. I just finished learning basics of python. I wanna build up my experience, so I hope I can work on real-world projects for free. How can I get that kind of opportunity?

LLuka_T Any tips for cartoon character lora training? Which caption approach and which c...

Furkan Gözükara SECourses•7/31/24, 11:09 AM

prepare a good dataset over 200 images

Furkan Gözükara SECourses•7/31/24, 11:09 AM

for captioning i usually test different captioners

Furkan Gözükara SECourses•7/31/24, 11:09 AM

you can locally use llava or cog vlm v1

JJP525 Hi everyone. I just finished learning basics of python. I wanna build up my expe...

Furkan Gözükara SECourses•7/31/24, 11:09 AM

you can watch channels and see if anyone post job offer

FFurkan Gözükara SECourses you can watch channels and see if anyone post job offer

JP525•7/31/24, 11:51 AM

I am not seek a job offer for money, I wanna learn from a seasoned developer while I am working for him for free.

JJP525 I am not seek a job offer for money, I wanna learn from a seasoned developer whi...

Furkan Gözükara SECourses•7/31/24, 12:17 PM

ok i hope someone in needs you message you

FFurkan Gözükara SECourses ok i hope someone in needs you message you

JP525•7/31/24, 12:37 PM

Thanks. Do you have anyone to recommend to me?

JJP525 Thanks. Do you have anyone to recommend to me?

Furkan Gözükara SECourses•7/31/24, 12:38 PM

not that i know atm

Automatic1111•7/31/24, 7:32 PM

@Furkan Gözükara SECourses hey my training finished in kohya but I didn't have enough space, is there a way to resume training?

Automatic1111•7/31/24, 7:32 PM

I didn't hav any saved epochs sadly

Automatic1111•7/31/24, 7:32 PM

I am furious

Automatic1111•7/31/24, 7:32 PM

It's for a LoRA

Automatic1111•7/31/24, 7:35 PM

It went for 18hrs :/

Automatic1111•7/31/24, 7:52 PM

Oh shoot it was a dreambooth model not a LoRA

AAutomatic1111 Oh shoot it was a dreambooth model not a LoRA

Furkan Gözükara SECourses•7/31/24, 8:15 PM

you cant resume sadly

Furkan Gözükara SECourses•7/31/24, 8:15 PM

if you have safetensors file you can use that as a resuming point

Furkan Gözükara SECourses•7/31/24, 8:15 PM

if you have any

Automatic1111•7/31/24, 8:23 PM

Sadly I don't but I'll just run it again. It's 3300 images (of my own artwork) and It'll be done 78 hours from now

AAutomatic1111 Sadly I don't but I'll just run it again. It's 3300 images (of my own artwork) a...

Furkan Gözükara SECourses•8/1/24, 1:02 AM

ye that is a big dataset

Furkan Gözükara SECourses•8/1/24, 1:02 AM

save frequent checkpoints

Automatic1111•8/1/24, 1:02 AM

I have another of about 14,000 images that I want to use also

Automatic1111•8/1/24, 1:02 AM

That's a big dataset?

AAutomatic1111 That's a big dataset?

Furkan Gözükara SECourses•8/1/24, 1:04 AM

not small one.

Furkan Gözükara SECourses•8/1/24, 1:04 AM

probably you need bigger vram and higher batch size

Automatic1111•8/1/24, 1:09 AM

I'm running a 3090 and had a look at your patreon and stuff

Automatic1111•8/1/24, 1:12 AM

I tried some of your settings but I had some errors for Kohya

AAutomatic1111 I tried some of your settings but I had some errors for Kohya

Furkan Gözükara SECourses•8/1/24, 2:21 AM

if you use single GPU just increase batch size, enable xformers and disable gradient checkpointing

Furkan Gözükara SECourses•8/1/24, 2:21 AM

all rest is same

Furkan Gözükara SECourses•8/1/24, 2:21 AM

multi gpu is harder though

Automatic1111•8/1/24, 2:23 AM

Should I make the batch size 2? 4?
Sorry SweatAnim

AAutomatic1111 Should I make the batch size 2? 4? Sorry <a:SweatAnim:746363251998064664>

Furkan Gözükara SECourses•8/1/24, 2:26 AM

maximum according to your GPU vram

Furkan Gözükara SECourses•8/1/24, 2:26 AM

if you need speed

Furkan Gözükara SECourses•8/1/24, 2:26 AM

but with 3090 you cant even get 2

Furkan Gözükara SECourses•8/1/24, 2:26 AM

maybe onetrainer can get

Furkan Gözükara SECourses•8/1/24, 2:26 AM

kohya cant

Furkan Gözükara SECourses•8/1/24, 2:27 AM

if you are using onetrainer you can rent A6000 on Massed Compute and do at least batch size 4

Xcorpion•8/1/24, 5:19 AM

do we have a queue in kohya ss?

Luka_TOP•8/1/24, 6:54 AM

When I install your caption models, where do they store? I tested some models and I don't like them so I want to get rid of Gigs and Gigs of models downloaded to my disk.

FFurkan Gözükara SECourses if you have safetensors file you can use that as a resuming point

Automatic1111•8/1/24, 12:23 PM

Alright, I think I have my first epoch, or I will later today.
Can you show me how to resume training?

I want to stop training so that I can make some sprite-style images, and then resume training later?

XXcorpion do we have a queue in kohya ss?

Furkan Gözükara SECourses•8/1/24, 1:03 PM

what do you mean by that?

LLuka_T When I install your caption models, where do they store? I tested some models an...

Furkan Gözükara SECourses•8/1/24, 1:03 PM

hugging face cache unless downloaded in local folder

Furkan Gözükara SECourses•8/1/24, 1:03 PM

check this article

Furkan Gözükara SECourses•8/1/24, 1:03 PM

https://www.patreon.com/posts/108419878

Patreon

How to Get Installation Logs and How to Change Hugging Face Cache F...

Get more from SECourses: Tutorials, Guides, Resources, Training, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözükara on Patreon

Furkan Gözükara SECourses•8/1/24, 1:04 PM

i chnaged my cache to this folder

AAutomatic1111 Alright, I think I have my first epoch, or I will later today. Can you show me h...

Furkan Gözükara SECourses•8/1/24, 1:04 PM

the resume means

Furkan Gözükara SECourses•8/1/24, 1:04 PM

you use your checkpoint.safetensors file

Furkan Gözükara SECourses•8/1/24, 1:04 PM

as a base model

FFurkan Gözükara SECourses as a base model

Automatic1111•8/1/24, 3:11 PM

And then I just train it with the same exact training data ?

AAutomatic1111 And then I just train it with the same exact training data ?

Furkan Gözükara SECourses•8/1/24, 3:39 PM

yes if you want to have effect of continue